American artificial intelligence company Anthropic announced that its large model assistant Claude now has the ability to directly control the user's computer and perform various operations on his behalf, further strengthening its "agent-type" functions to cope with the competitive pressure brought by the open source agent framework OpenClaw this year.
According to Anthropic's announcement, as long as users subscribe to an eligible paid plan (such as Claude Pro or Claude Max), they can authorize Claude to take over the computer and complete specific tasks such as finding and sending files from the local hard drive. Behind this new feature is the overall rise in agent-based AI capabilities in the industry. Among them, the open source framework OpenClaw, which became popular earlier this year, has spawned a complete ecosystem. Many tools called "claws" can perform operations relatively autonomously on the user's computer or connected system based on simple instructions. Last week, NVIDIA released the NemoClaw framework, which aims to simplify the deployment and installation of OpenClaw agents, while adding certain security settings to further promote the development of this ecosystem.
Anthropic said that after receiving a task, Claude will prioritize finding suitable tools through "connectors" to various applications, such as integrating services such as Google Calendar or Slack. If relevant tools or connectors are not available, Claude can also directly simulate user operations and perform tasks through keyboard input and mouse movement, including using browsers, development tools, and opening local files. During the execution process, Claude will scroll, click, and other operations like an ordinary user, but the user's permission is required before each actual operation, and the user can terminate Claude's behavior at any time.
However, although letting chatbots control computers is attractive in terms of efficiency, it also raises security and risk control concerns. Experts point out that a major risk of so-called "agent AI" is that it can perform large-scale, potentially far-reaching operations in a very short time, and users may not be aware of it in time. In addition, if "claw" tools such as OpenClaw are hijacked by malicious attackers, they may use users' personal data and systems to perform unauthorized actions. In response to these risks, Anthropic said that it has added multiple security mechanisms to the new features to minimize the threats of attack methods such as "prompt injection". The system will automatically scan for related risks and continue to update protection when new vulnerabilities appear.
Despite this, Anthropic still issued a clear warning to users on the product page, emphasizing that this computer control function is still in the research preview stage and may have errors or abnormal behavior. It is recommended that users avoid using it for applications that handle highly sensitive data. In fact, the company has disabled some apps involving sensitive information by default to reduce the risk of misuse. Currently, the research preview of this feature is only open to macOS computers, and is only available for trial use to Claude Pro and Claude Max subscribers.
Anthropic also emphasized that the combination of this computer control capability and its "Dispatch" function will significantly expand Claude's application scenarios. Dispatch allows users to assign tasks to Claude through terminals such as mobile phones, such as checking email every morning, or starting specific work sessions such as Claude Cowork and Claude Code. When computer control and Dispatch are enabled at the same time, Claude can complete more automated processes without the user being in front of the computer, such as generating morning information briefings, running test tasks, etc.
Given that computer control capabilities and Dispatch are both newly launched features, Anthropic also admits that some complex tasks may not be completed successfully at the current stage. The company stated that the reason for opening it in advance in the form of a "research preview" is to obtain feedback from real users as early as possible and understand where the system needs to be improved most in actual use, so as to polish Claude into a more powerful and reliable productivity tool.